Better Tropes
A catalogue of the tells and styles that mark writing as AI-generated, with a plain rewrite for each.
What is Better Tropes?
Better Tropes is a catalogue of 49 tells that make writing read as AI-generated, grouped into 7 categories, each with a clear alternative to use instead. It's a Markdown file. Paste it into a system prompt, download it, or install it as an agent skill.
# Better Tropes
These tells make writing sound AI-generated. Avoid them while following your other instructions. Each entry records a tell, and explains it alongside a pair of examples. Always prefer the second:
- **✗** the tell to avoid
- **✓** the rewrite to prefer
---
## Word Choice
### Elevated Filler Verbs
Verbs that sound important but carry no more meaning than a plain one: `delve`, `leverage`, `utilize`, `harness`, `foster`, `streamline`, `unlock`, `embark`, `spearhead`, `navigate`. Prefer the shortest verb that is true.
- **✗** We leverage data to unlock growth and streamline collaboration.
- **✓** We use data to grow and to help teams work together.
### Magic Adverbs
Adverbs used to make a flat statement feel weighty: `quietly`, `deeply`, `fundamentally`, `remarkably`, `arguably`, `seamlessly`, `effortlessly`, `truly`, `simply`. If the sentence means the same without the adverb, cut it.
- **✗** The system quietly and fundamentally reshapes how teams work.
- **✓** The system changes how teams assign work.
### Throat-Clearing
Filler that announces significance instead of showing it: `it's worth noting`, `it bears mentioning`, `importantly`, `interestingly`, `notably`, `certainly`. State the thing plainly.
- **✗** It's worth noting that, importantly, retention improved.
- **✓** Retention improved 12%.
### Abstraction Nouns
Words that gesture at scale or at a vague boundary instead of naming a thing: `tapestry`, `landscape`, `realm`, `journey`, `ecosystem`, `paradigm`, `synergy`, `framework`, `seam`, `seams`, `world of`. Name the concrete noun instead.
- **✗** Navigating the ever-evolving landscape of the customer journey.
- **✓** Keeping up with what customers ask for.
### Padding Adjectives
Adjectives that add length, not information: `robust`, `comprehensive`, `holistic`, `seamless`, `powerful`, `innovative`, `structural`. Delete them or replace them with a specific.
- **✗** A robust, comprehensive solution with a structural advantage.
- **✓** It handles 10,000 requests a second and never drops one.
### Empty Superlatives
Claims of importance with nothing behind them: `game-changing`, `cutting-edge`, `state-of-the-art`, `revolutionary`, `best-in-class`. Show the thing; let the reader judge.
- **✗** Our game-changing, cutting-edge platform.
- **✓** Our platform ships changes in under a minute.
### The Long Word for a Short One
The longer, Latin-rooted word where a short one works: `utilize` (use), `facilitate` (help), `commence` (start), `endeavour` (try), `demonstrate` (show).
- **✗** We endeavour to facilitate a commencement of the process.
- **✓** We try to help you get started.
### The "Serves As" Dodge
Weak linking verbs that assert importance without stating a fact: `serves as`, `stands as`, `marks`, `represents`. Say what the thing actually does.
- **✗** This release serves as a testament to our commitment.
- **✓** This release cuts load times in half.
---
## Sentence Structure
### The "It's Not X - It's Y" Antithesis
The single most-recognised tell: a negation followed by an elevated correction, often across an em dash, and its cousin `not because X, but because Y`. Say the positive thing directly.
- **✗** It's not just a tool - it's a way of thinking.
- **✓** The tool changes how you plan your week.
### "Not X. Not Y. Just Z."
Staccato negation fragments building to a reveal: `Not a bug. Not a feature. Just gone.` Write the point as a sentence.
- **✗** Not a tweak. Not an update. A rebuild.
- **✓** We rebuilt it from scratch.
### The Rhetorical Fragment
A question-then-fragment beat for drama: `The result? Devastating.` / `The catch? There isn't one.` Answer in a full sentence.
- **✗** The result? A total transformation.
- **✓** Support tickets dropped by half.
### Repeated Sentence Openings
Consecutive sentences opening with the same words for cadence: `They assume... They expect... They forget...`. Use it once for real effect; delete the rest.
- **✗** They want speed. They want scale. They want simplicity.
- **✓** They want it fast, and they want it simple.
### The Rule-of-Three Cadence
Three parallel items in every sentence, whether or not three things are true. Vary the count; two is often enough, and one is often better.
- **✗** Fast, reliable, and scalable - built to grow, adapt, and thrive.
- **✓** It's fast, and it holds up under load.
### "Not Only... But Also"
A correlative frame that inflates a simple point into a two-part announcement. Split it or drop it.
- **✗** This not only saves time but also improves accuracy.
- **✓** This saves time and cuts errors.
### The Trailing "-ing" Clause
A sentence that trails off into an `-ing` clause explaining its own significance: `..., ensuring...`, `..., allowing...`, `..., highlighting its importance`, `..., reflecting broader trends`, `..., contributing to`. End the sentence at the point.
- **✗** The API caches results, ensuring a seamless experience and reflecting broader trends.
- **✓** The API caches results, so pages load instantly.
### Uniform Rhythm
Every sentence the same medium length, so the prose has no pulse. Vary sentence length. Cut some short. Let one run long.
- **✗** The feature is useful. It saves time. Users like it. It works well.
- **✓** The feature saves time. Not a little - users get their afternoons back.
### The "From X to Y" Sweep
A description of range in place of specifics: `from startups to enterprises`, `spanning every use case`. State the examples instead of describing the range.
- **✗** From onboarding to offboarding, spanning teams of every size.
- **✓** We handle the first-day setup and the account-deletion request.
---
## Paragraph Structure
### The Restated Heading
A paragraph whose first sentence restates the heading in fuller words before saying anything new. Delete the restatement and open on the real point.
- **✗** _(under "Pricing")_ When it comes to pricing, there are several factors to consider.
- **✓** _(under "Pricing")_ Most teams pay about $40 a seat.
### The Preview-Recap Sandwich
Signposting that narrates the structure instead of delivering it: `In this section, we'll explore...`, `As we've seen...`. Cut the narration and start.
- **✗** In this section, we will explore three ways to reduce churn.
- **✓** Three things reduce churn. The first is onboarding.
### Listicle in a Trench Coat
Prose that is secretly a list, stitched together with `The first...`, `The second...`, `The third...`, or bulleted for the feel of being organised. Use a list for genuinely parallel items; use sentences for an argument.
- **✗** The first reason is speed. The second reason is cost. The third reason is trust.
- **✓** It's faster and cheaper, and most importantly - people trust it.
---
## Tone
### Relentless Positivity
A pitch with no downside, no cost, no trade-off. Real recommendations are often granular, tangible, spiky, or include drawbacks.
- **✗** This approach is a win for everyone involved.
- **✓** This is faster to ship, but you'll pay more in support tickets.
### Manufactured Excitement
Manufactured excitement: `thrilled`, `excited to share`, `can't wait`, and the exclamation marks that follow. Let the news carry the feeling.
- **✗** We're thrilled to announce our exciting new feature!
- **✓** You can now schedule posts a week ahead.
### The "Let's / Here's" Warm-Up
Fake conversational lead-ins aimed at no one: `Let's dive in`, `Let's break this down`, `Let's unpack this`, `Let's explore`, `you might be wondering`, and the `Here's the ___` slot filled with almost any noun. For example - `Here's the thing`, `the kicker`, `the rub`, `the secret`, `the takeaway`, `the lesson`, `the deal`, `the starting point` - plus `Here's where it gets interesting` and `Here's what most people miss`. Address the reader like a busy adult and start.
- **✗** Let's break this down. Here's the kicker: it's simpler than you think.
- **✓** Here is how it works.
### "Think of It As..."
Reaching for an analogy the moment a plain description would do: `Think of it as`, `Think of it like`, `It's like a`. Describe the actual thing first.
- **✗** Think of it as a Swiss Army knife for your inbox.
- **✓** It sorts, snoozes, and schedules email in one place.
### "Imagine a World Where..."
Opening on a hypothetical instead of the present fact: `Imagine a world where`, `Picture this`. Start with what is true now.
- **✗** Imagine a world where your reports write themselves.
- **✓** Your weekly report now drafts itself.
### False Vulnerability
Simulated self-awareness staged for warmth: `I'll admit, I'm obsessed with this`, `Okay, confession time`. If there's no real admission, don't perform one.
- **✗** And yes, I'm openly in love with this feature.
- **✓** This is the feature I use most.
### "The Truth Is Simple"
Claiming hard-won clarity to sound authoritative: `The reality is simpler`, `History is unambiguous`, `The real story is`. State the point without the frame.
- **✗** The reality is simpler than the experts would have you believe.
- **✓** Cheaper shipping raised conversions. That's the whole story.
### The Fence-Sitter
Refusing to hold a position, presenting every side as equally valid so nothing is actually said. If the piece has a view, state it.
- **✗** There are many perspectives, and each has its merits.
- **✓** Monthly billing is the wrong default. Here's why.
### Grandiose Stakes Inflation
Inflating a modest thing into an epoch: `fundamentally reshape`, `define the next era`, `something entirely new`, `change everything`. Match the claim to the size of the thing.
- **✗** This feature will fundamentally reshape how humanity works.
- **✓** This feature saves your team an hour a week.
### Invented Concept Labels
Coining an official-sounding term for an ordinary idea: `the supervision paradox`, `the acceleration trap`, `workload creep`. Describe the thing plainly instead of branding it.
- **✗** This is what I call the productivity paradox.
- **✓** The more tools people add, the less they get done.
### The Motivational Closer
An ending that swells into uplift instead of landing the point: `The future is bright`, `The possibilities are endless`, `The journey has just begun`. End on something concrete.
- **✗** The possibilities are truly endless.
- **✓** Try it on one project this week and see.
---
## Formatting
### Em-Dash Overuse
The em dash - or the double-hyphen `--` standing in for one -
as a default connector several times a paragraph. Reach for a full stop, a comma, or parentheses first; keep the dash for when nothing else works.
- **✗** It's fast - really fast - and it scales - effortlessly.
- **✓** It's fast, and it scales.
### Bold Everything
Bolding so many phrases that emphasis stops meaning anything. Bold at most the one thing a skimmer must not miss.
- **✗** **Speed**, **reliability**, and **scale** are **critical** for **success**.
- **✓** **Speed** matters most. Everything else is negotiable.
### Bold-First Bullets
Every bullet formatted as `**Bolded phrase**: description`, turning a list into a glossary. Use it when terms genuinely need defining, not as the default bullet shape.
- **✗** **Speed**: it's fast. **Value**: it's cheap. **Trust**: people like it.
- **✓** It's fast, it's cheap, and people trust it.
### Unicode Decoration
Decorative characters standing in for structure: arrows (`→`), curly "smart" quotes where straight ones read fine, and stray symbols. Keep punctuation plain unless the format calls for more.
- **✗** Fast → Reliable → Scalable ✦
- **✓** Fast, reliable, and built to scale.
### Emoji Section Markers
Decorative emoji prefixing headings and bullets (🚀, ✅, 💡) as a substitute for structure. Remove them unless the format genuinely calls for them.
- **✗** 🚀 Getting Started · ✅ Key Benefits · 💡 Pro Tips
- **✓** Getting started. What you get. One thing to try.
---
## Composition
### The Definition Opener
Starting on an encyclopedia-style definition of the topic instead of the point: `X is a Y that...`. Open where the reader's interest is, not at the dictionary.
- **✗** Customer churn is the rate at which customers stop using a product.
- **✓** You lost 40 customers last month. Here's where they went.
### Scope Inflation
Opening on a grand backdrop to make a small point feel large: `In today's fast-paced digital world`, `In an era of rapid change`. Start on the specific.
- **✗** In today's fast-paced digital landscape, email still matters.
- **✓** Your open rate dropped 8% last quarter.
### The Unnamed Source
Authority attributed to no one: `Experts argue`, `Studies show`, `Industry reports suggest`, `Observers have cited`. Name the source or drop the claim.
- **✗** Studies show that most users prefer dark mode.
- **✓** In our 2026 survey of 3,000 users, 61% chose dark mode.
### Historical Analogy Stacking
Rapid-fire comparisons to famous companies or moments to borrow their weight: `the Uber of...`, `like Apple in 2007`, `a Spotify moment`. Make the case on its own merits.
- **✗** It's the Uber of groceries, with an Apple-grade design and a Netflix model.
- **✓** You order in two clicks and it arrives in an hour.
### The Token Caveat Ending
Conceding a token challenge only to reaffirm the upside: `While challenges remain, the outlook is positive`, `Despite its hurdles...`. If there's a real challenge, take it seriously. If not, don't stage one.
- **✗** While hurdles remain, the future looks bright.
- **✓** The hard part is migration. Budget two weeks for it.
### The Signposted Conclusion
A closing paragraph flagged as one - `In conclusion`, `To sum up`, `In summary` - that restates the piece without adding anything.
- **✗** In conclusion, this is a powerful approach that can transform your workflow.
- **✓** If you remember one thing: chunk the work before you prompt.
### Restating the Prompt
Answering by echoing the question back in full-sentence form before responding.
- **✗** That's a great question. The best way to reduce load times is...
- **✓** Serve images at the size they display. That's most of it.
---
## Whole-Piece Patterns
These signs of LLM writing only become visible once the piece is finished. Check for them on a read-through, not a find-and-replace.
### One-Line Paragraphs
Breaking a thought into one-sentence paragraphs, each on its own line, to force drama. And it shows. Every time. Reserve the isolated line for the one point that earns it.
- **✗** Three consecutive one-line paragraphs, each "landing" a point.
- **✓** A normal paragraph, then one short line where it actually matters.
### Symmetrical Paragraphs
Every paragraph the same shape and length, like slots in a template. Let content decide length; a one-line paragraph is allowed to stand alone.
- **✗** Three paragraphs, each exactly four sentences, each opening with a thesis.
- **✓** A five-sentence paragraph, then a two-word one. It matters.
### Heading Inflation
Turning every second sentence into a Title-Cased heading, breaking a short piece into a dozen sections. Use a heading when the reader needs to jump; otherwise let prose run.
- **✗** Six headings over 300 words.
- **✓** One heading, three paragraphs under it.
### The Overworked Metaphor
A single figure of speech introduced early and then dragged through the whole piece - every point reframed as another part of the "journey", "engine", or "recipe". Use a metaphor once, then let it go.
- **✗** _A "building a house" metaphor reused in all six sections._
- **✓** _One clear image where it helps, then plain language after._
### One-Point Dilution
A single idea restated across thousands of words, each section rephrasing the last. Make the point once, support it, and stop.
- **✗** _Five sections that all say "clarity beats cleverness" in new words._
- **✓** _One section that says it, then an example that proves it._
### Content Duplication
The same sentence or paragraph reappearing near-verbatim within one piece. Read it through once and cut every echo.
- **✗** The intro claim repeated word-for-word in the conclusion.
- **✓** The conclusion adds something the intro didn't.
---
_Better Tropes — maintained at <https://fromfireside.com/tools/better-tropes> · source: <https://github.com/fromfireside/better-tropes> · CC BY 4.0_
What's in the catalogue
49 entries across 7 categories. Each one links into the file above.
Who makes Better Tropes, and why?
I'm Sean O'Neill, the founder of Fireside, and I help organisations tell their story with automated content. I needed a prompt file that I could use with my clients. tropes.fyi is a good catalogue, but it is written for a human audience; the comments that make it readable add noise to system prompts. Better Tropes is a simpler one.
A lot of prompt files conflate "reducing AI tells" with "improving writing style". They are separate tasks. When you produce large volumes of content for organisations, your writing style prompts should be bespoke to that organisation - you can't use something from the internet for this. As a result, I use Better Tropes to avoid AI tells, and I handle the content style myself. That's why Better Tropes exists, and that's how you should use it.
It's in a mature state, but I add to it from time to time.
Why use Better Tropes instead of a list of AI words to avoid?
A word list tells a model what to avoid but doesn't improve overall writing style. The writing will still sound like slop.
Of the 49 entries here, 8 are word choice. The other 41 are sentence structure, paragraph structure, tone, formatting, composition and whole-piece patterns: the rule-of-three cadence, the signposted conclusion, bold-first bullets, the preview-recap sandwich. Words alone do not mean that text is AI-generated slop.
Everything here is something you can point at in a finished draft. It is not advice on how to write. Used as is, it cuts how often the tells appear, often by 30-70% in my testing. Simpler tasks perform better.
How is Better Tropes different from tropes.fyi, unslop and humanizer?
Better Tropes only lists AI tells. It has no instructions on how to write and no commentary on why a tell happens, so everything you paste is a rule with an alternative. It lists 49 tells in 6,047 tokens. tropes.fyi, the other prompt file here without writing advice, lists 49 in 9,505.
tropes.fyi explains each tell to a person reading about AI writing and tracks whether it is rising or fading, and its prompt file carries that commentary with it. Unslop, stop-slop, humanizer and avoid-ai-writing are agent skills that pair a list of tells with rules on how to write, such as using the active voice or matching a writer's voice. Wikipedia's Signs of AI writing is a guide for editors reviewing articles.
| Resource | Best for | Tells | Tokens | Writing advice | License |
|---|---|---|---|---|---|
| Better Tropes | A tells-only rule set beside your own style prompts | 49 | 6,047 | No | CC BY 4.0 |
| tropes.fyi | Learning why each tell reads as AI | 49 | 9,505 | No, but commentary on every entry | Not stated |
| avoid-ai-writing | The most thorough audit of a finished draft | 74 | 46,664 | Yes | MIT |
| humanizer | Rewriting in a particular writer's voice | 26 | 10,896 | Yes | MIT |
| stop-slop | A short skill with a firm house style | 19 | 4,888 | Yes | MIT |
| unslop (Cursor pstack) | The smallest file to leave on in a coding agent | 28 | 2,251 | Yes | MIT |
| Wikipedia: Signs of AI writing | Reviewing text someone else wrote | Not counted | Web page | No | CC BY-SA 4.0 |
Which skill should I use to remove AI writing tells?
For the most thorough check of a finished draft, use avoid-ai-writing. It covers 74 patterns, grades each finding by severity and can flag problems without rewriting. It is also by far the largest file here, so it suits an editing pass better than a standing system prompt.
To rewrite text so it sounds like a particular person, use humanizer. Give it a sample of their writing and it matches how they write.
For the smallest file to leave switched on in Cursor or another coding agent, use unslop from Cursor's pstack plugin. At about 2,300 tokens it covers the common tells and adds a few plain-writing rules.
If you want a short skill that also enforces a style, such as active voice and no adverbs, use stop-slop. Its style rules are firm, so check they suit your writing first.
To understand why a pattern reads as AI, or to teach an editor to spot it, read tropes.fyi. If you are reviewing text someone else wrote, Wikipedia's Signs of AI writing has real examples from flagged articles.
Use Better Tropes when you already write your own style prompts and want the tells gone, without a second set of style opinions competing with yours.
How do you use Better Tropes?
Paste the whole file into a system prompt, a project instruction or a custom GPT as part of your instructions. It works best as a standing constraint rather than a one-off editing pass. Once you have your other prompts, data, style, output, channel etc dialled in - then revisit Better Tropes and shorten it by removing descriptions and unnecessary examples that would never appear in your task. I've kept noise low, but the simple act of having examples and instructions clutters up a prompt file; once you know what you're doing, go back and remove these for improved results.
Download the Markdown if you would rather keep it in a repo, or
install it as an agent skill so your agent loads it when it writes.
The skill is the same catalogue with frontmatter on it. If you have
multiple projects, I recommend renaming each one as a unique
instance, e.g. my-project-better-tropes. That way you
can tailor each copy to its own task; one generic version has to
serve everything.
For the best results, cut it down to suit your workflow. A technical blog post needs different entries from a social post. I've written about tailoring in how to reduce LLM hallucinations.
How do you make AI text sound less like slop?
Provide better data to the model. Write your prompts by hand. If all your prompts are written by LLM, of course your text sounds robotic. Models do not act on "write better" or "sound more human", but they DO mirror prompt style *a lot*. Make your instructions very explicit, and quantify or supply data to support the output you want.
For a draft you've already written, hand this file to Claude or ChatGPT and ask it to de-slop the text. Both do a decent job of it.
Which words does AI overuse most?
Everyone knows delve, span,
tapestry and robust. But swap
delve for dig and the text is still
obviously AI. It's the recurring combination of patterns that gives
it away. Word choice varies far more in human writing than people
realise.
The word-level tells are grouped into Elevated Filler Verbs, Abstraction Nouns, Padding Adjectives and Empty Superlatives, each with the preferred alternative to use instead.
Why does AI writing sound like that?
Training rewards answers that sound agreeable and complete, so models hedge, qualify and inflate. That is where relentless positivity, the manufactured excitement and the closing paragraph that announces its own significance come from.
Instruction-following does the rest. A model asked for a structured piece will restate the heading in its first line, preview what is coming and recap it at the end. The markers of structure are easier to produce than the structure itself.
Does using Better Tropes, or removing AI tells, beat an AI detector?
No. AI text detectors measure predictability and burstiness. Better
Tropes helps you reduce slop and get text that sounds more human; it
doesn't increase text variability per se. Taking
delve out of a draft does not move the "is it AI"
number much, but it's much nicer to read.
If you are trying to pass an academic integrity check or other scorecard, this is the wrong tool.
Can Better Tropes reduce hallucinations?
No. Better Tropes is only for making text sound less AI and less like slop. It may indirectly reduce hallucinations, because a lot of Better Tropes patterns are about being short and succinct. A model that is being too verbose is more likely to reach for facts either from its training data or invent some, but Better Tropes is not the right tool if your primary aim is reducing hallucinations.
To reduce hallucinations and improve the factual accuracy of your output, see: How to Reduce LLM Hallucinations and Keep Writing Human.
Inspired by tropes.fyi - who taught us to expect better - and our own experience operating Fireside. Its companion for writing about people rather than product is Better Recognition.
From the inbox
"loving better tropes, combining it with ste100 in just about everything I touch these days as a global output style. God I hate claude's default word salad."
"This is awesome. VERY nice"